I'd say it's neither. The only failures I've seen in the data center were caused by differential thermal expansion cycles and broken solder balls. Same phenomena that kills game consoles but not ML/mining GPUs that spend all day and night at max power/current and constant temperature.
Until you power cycle them a couple of times... My old AMD GPU even held a record for how "good" the ASIC still was or was not.
TL;DR Yes, your GPU wears and becomes slower over time; eventually it brakes.
Not because of bad thermal coefficients but because of the current... Even faster so when OC'ing because U=I*R. You can benchmark it yourself.
The broken solder balls of the past where attributed to the transition from leaded to lead free solder. If I recall correctly, it was mostly NVIDIA and Apple who suffered from this and only temporarily / 1 or 2 generations.